Lech Kalinowski is a Senior AI System Engineer at Callstack, a physicist specializing in correlated electronic systems, and a co-founder of the sustainable-electronics startup Panamint. His work spans printed sensors, mobile inference, coding-agent evaluation, and experimental hardware for interacting with self-hosted AI.
Kalinowski studied at the University of Silesia in Katowice and earned a doctorate in physics. He previously served as chief technology officer at GamerHash before co-founding Panamint, where he also became CTO.
A conversation at the 2019 InterNano conference in Katowice led Kalinowski, Mateusz Bryning, and Adam Szatkowski to explore commercially viable sustainable electronics. They founded Panamint in 2020, initially developing printed sensors from graphite and graphene flakes before expanding into smart labels and systems for monitoring sensitive goods. The company’s founding story traces Kalinowski’s progression from electronic materials to practical hardware development.
Building AI for real-world constraints
- Vault: a handheld terminal for self-hosted agents. Kalinowski built Vault, an ESP32-based device for controlling OpenClaw agents. Its OLED screen handles live interaction, while electronic paper preserves longer responses. Fixed buffers, preallocated memory, and one-bit graphics keep the firmware lightweight; a separate NVIDIA DGX Spark runs model inference through TensorRT-LLM. Kalinowski also developed a text-based role-playing mode that generates characters and settings while preserving narrative context. His AI Engineer demonstration illustrates the deliberate contrast between a quiet, low-power handheld and the powerful computer supporting it.
- Task-specific local inference. Kalinowski recommends on-device models for short, private, latency-sensitive tasks, while reserving demanding reasoning and repository-scale coding for stronger remote systems. He has also built an offline mobile image-generation system guided by ambient-light sensors and accelerated by a neural processing unit. His writing on quantization treats model precision as a practical trade-off between memory, latency, cost, and output quality.
- Structured context before unnecessary retrieval. For bounded, relatively stable knowledge problems, Kalinowski favors structured context injection, deterministic preprocessing, and schema constraints over elaborate retrieval infrastructure. He reserves retrieval for workloads where scale, changing documents, citation requirements, or multitenancy justify its complexity.
- React Native Evals and agent coordination. Kalinowski helped develop React Native Evals, an open-source benchmark covering practical React Native coding tasks, and co-developed an evaluation workflow using Airflow and MLflow. He also created Agents Commander, a lightweight interface for multi-agent command-line workflows, and investigated how independently running small models can cooperate without obscuring their actual infrastructure costs.
